DatarrX/myanmar-synthetic-syllable-glyphs

🇲🇲 Myanmar Synthetic Syllable Glyphs (MSSG)

The Myanmar Synthetic Syllable Glyphs (MSSG) is a massive-scale, high-fidelity synthetic image dataset containing 14,295,552 heavily augmented glyph images (128x64 pixels, grayscale) representing the structural combinatorial matrix of the Burmese script.

Developed and engineered by Khant Sint Heinn (Kalix Louis), this core foundational dataset is officially published and maintained under DatarrX (Myanmar Open Source Organization, NPO) to advance low-resource Natural Language Processing (NLP), Optical Character Recognition (OCR), and Handwritten Text Recognition (HTR) frameworks for the Myanmar language.

---

📌 Project Motivation & Core Concept

While standard datasets focus strictly on common dictionary words, modern deep learning architectures (like ViT-based OCRs or Vision-Language Models) require deep structural familiarity with character bounds and extreme visual variations to achieve near-perfect accuracy.

To solve this, MSSG bypasses the limitations of traditional corpora by treating the Myanmar language as a generative combinatorial grid. By creating an automated pipeline that maps out the extensive "အ Pattern" rules and systematically substituting the 36 primary graphemes (the 33 traditional Burmese consonants plus ဉ, ဥ, and ဧ), the project captures an exhaustive evolutionary blueprint of native Burmese orthography.

---

⚠️ Important Scientific Note: Combinatorial Coverage

A crucial and deliberate design aspect of the MSSG dataset is the inclusion of non-standard, rare, and historically unutilized syllable structures. Because the generation pipeline mathematically iterates through every theoretical permutation of consonants, vowel markers, medials, and diacritics, it naturally produces sound-glyph combinations that do not actively exist in contemporary spoken or written Burmese.

Why include non-existent syllables?

1. Structural Robustness: Training vision models on these valid but unutilized physical shapes forces the neural network to master the actual engineering of the Burmese stroke boundaries rather than relying on context clues or linguistic biases.
2. Historical & Computational Completeness: It provides zero-shot compatibility for ancient scripts, archaic spellings, regional dialects, or rare loan words that conventional modern dictionaries entirely omit.

---

🛠️ Generative Patterns & Grapheme Rules

The entire generative architecture is built upon a curated list of base patterns that model how the Burmese script wraps around an anchor character.

The structural foundations generated within this dataset include the fundamental bare-vowel markers, compound vowel configurations, complex diacritic stacking, lower-element subscript ligatures (ဗျည်းဆင့်), horizontal and vertical medial combinations, heavy terminal glottal stops, specialized nasalizations, and multi-layered multi-syllabic glyph clusters. Additionally, a distinct subset of autonomous special characters—including sacred historical vowels, archaic punctuation symbols, abbreviation markers, standard numerals, and compound ligatures—is generated to ensure absolute linguistic coverage.

To expand this matrix into a realistic dataset, strict substitution and phonological extension rules were executed programmatically:

• The Medial Expansion Rule: For every base pattern featuring the medial-ya (ျ), a parallel equivalent is dynamically spawned substituting the medial-ra (ြ).
• The Vowel Alternation Rule: For every pattern containing the standard long-a vowel sign (ာ), an alternate variation is generated utilizing the tall-a sign (ါ) to accurately represent native orthographic variations.

---

📊 Dataset Specifications & Features

• Total Row Count: 14,295,552 instances
• Total Footprint: ~10.2 GB (Compressed Parquet IPC Stream)
• Image Dimensions: 128x64 pixels (Optimized canvas to ensure tall stacked ligatures and deep subscripts are safely preserved without clipping)
• Color Space: 8-bit Grayscale (L mode), high-contrast white foreground stroke on a pure black background.
• Data Augmentations Applied: To prevent model overfitting, every unique string is passed through a four-stage variation pipeline: the clean pristine copy (original), controlled geometric rotational tilts (aug_1), horizontal/vertical pixel-shifting (aug_2), and micro Gaussian ink-bleeding blur simulations (aug_3).

---

👥 Roles & Credits

This project represents a monumentally scaled computational effort to democratize AI training blocks for the Myanmar community.

• Dataset Creator & Lead Architect: Khant Sint Heinn * Engineered the dynamic combinatorial iteration logic, structural padding matrix, automated rule expansion scripts, and the low-overhead memory-safe local Parquet packaging pipeline.
• Publisher / Supporting Organization: DatarrX * A dedicated Myanmar Open Source Non-Profit Organization fueling the development of free, accessible, and high-performance language assets for native technological independence.

---

📜 Citation & Academic Reference

If you incorporate the MSSG dataset into your academic research, industrial OCR pipelines, font-generation experiments, or language modeling benchmarks, please acknowledge the author and publisher via the following official BibTeX citation:

@misc{datarr_mssg_2026,
  author       = {Khant Sint Heinn},
  title        = {Myanmar Synthetic Syllable Glyphs (MSSG): A Large-Scale Combinatorial Grapheme Dataset for Structural OCR/HTR Training},
  year         = {2026},
  publisher    = {Hugging Face},
  organization = {DatarrX Initiative},
  howpublished = {https://huggingface.co/datasets/DatarrX/myanmar-synthetic-syllable-glyphs},
  note         = {Published under DatarrX Initiative}
}